Image OCR

Only after you first click 'start recognition' does the page load the recognition engine and language models (about 8.5MB, served entirely from this site's own domain, never through any third party); recognition then runs in a background thread of the browser - images never leave your machine and pass through no server. Verifiable in the Network panel: before the click there are no model requests at all, and during recognition there are no upload requests at all.

Simplified Chinese plus English (chi_sim + eng, the default) and English-only (eng, a smaller model and faster). Accuracy drops noticeably on handwriting, heavy blur, artistic lettering and low-contrast images.

On clean printed text both Chinese and English are usable, but expect occasional wrong characters (look-alike glyphs and line-break splits are the most error-prone); the page reports an average confidence and the result is editable - proofread by a human before formal use, which is the page's stated policy.

The first run downloads and initializes the engine and models (about 8.5MB); later runs reuse them. If the network is bad, refresh and retry. Models load on demand - until you press the button, this page requests not a single byte of any model.

This page only accepts images. For PDFs, first convert to JPG / PNG with our 'PDF to image' tool, then recognize; scanned PDFs lack a text layer, and the same image-based flow applies to them.